Kush_EXE

Can your boss tell ChatGPT or Claude wrote your email?

· Runs 2:49

On October 5, OpenAI said it will hide an invisible watermark in some ChatGPT writing, and Anthropic announced the same for Claude in August. It is not a "made with AI" stamp anyone can scan: the model nudges which words it picks, leaving a pattern only the company's secret key can find. Both companies say the mark carries no name, no organization and no chat.

So who can actually see it? Not you, for now. OpenAI's detector goes only to approved researchers and expert organizations, and checking student essays is outside its program. Anthropic's detector is a private preview for eligible groups such as regulators, the press, researchers and educational organizations. The twist for Americans: ChatGPT text gets the mark only in the EU, while newer Claude models mark text everywhere. We also cover how easily edits weaken the mark, why a missing mark proves nothing, how this differs from classroom AI checkers, and what to do if you get flagged. Not legal advice.

Watch on YouTube Subscribe

In this video

  1. Your AI writing may carry a hidden mark
  2. How it works, and can it be traced to you?
  3. Who can actually check it
  4. The twist for Americans: Claude vs ChatGPT
  5. Classroom AI checkers are a different thing
  6. Small edits weaken OpenAI's mark
  7. A missing mark proves nothing
  8. Proofreading vs translation
  9. Know the rules, and what to do if flagged

Your AI writing may carry a hidden mark 0:00

If AI wrote your cover letter, it might now carry a hidden watermark. You'd never see it. So can your boss spot it? And can it be traced back to you? On October 5th, OpenAI said it will hide an invisible watermark in some ChatGPT writing. Anthropic announced the same for Claude in August. You might picture a "made with AI" stamp anyone can scan. Let's check.

How it works, and can it be traced to you? 0:23

OpenAI says: no hidden characters. ChatGPT nudges which words it picks, leaving a pattern only a secret key can find. Can it be traced to you? Both companies say no: no name, no organization, no chat.

Who can actually check it 0:37

Can you check it yourself? No. For now, OpenAI's detector goes only to approved researchers and expert organizations, case by case. Anthropic's is a private preview for groups like regulators, the press and researchers. What about your teacher? OpenAI says checking student essays is outside its program, for now. But Anthropic's private preview does list educational organizations, among groups eligible under EU law.

The twist for Americans: Claude vs ChatGPT 1:03

Here's the twist for Americans. ChatGPT text gets the mark only in the European Union, because EU law requires it. Newer Claude models mark text everywhere, the US included, because Anthropic can't yet limit the mark by region. So between these two, in America it's Claude's writing, not ChatGPT's, that may already carry the mark.

Classroom AI checkers are a different thing 1:24

And classroom AI checkers? Different thing: without the key, they guess from writing style. Guessing goes wrong. In a 2023 Stanford study, seven detectors flagged real test essays by non-native English speakers as AI 61 percent of the time, on average.

Small edits weaken OpenAI's mark 1:41

OpenAI's watermark is fragile too. On passages of about 300 words, swapping one word in ten for a synonym cut detection from about 92 percent to 66. Swap one word in four: down to 17 percent. And under EU rules, texts under about 150 words don't even have to carry it.

A missing mark proves nothing 2:00

So a missing mark proves nothing. As OpenAI puts it, "The absence of a detected watermark does not prove human authorship." Even a hit can be wrong. False alarms are one reason OpenAI isn't releasing its detector publicly, at least at launch.

Proofreading vs translation 2:15

If Claude only proofreads your writing, there's little to mark. But a translation by Claude is fully marked. Anthropic says light edits probably won't fully scrub Claude's mark. But that misses the point: the real risk is breaking your school's or employer's AI rules.

Know the rules, and what to do if flagged 2:31

So know those rules and keep your drafts. If a checker flags you, ask which tool, and what evidence exists beyond the score. Verdict: this mark can't name you, and almost nobody can read it. So stop asking "will I get caught?" and start asking "am I allowed?"

What the sources say

  • OpenAI announced on 5 October 2026 that over the coming weeks it would add an invisible watermark to eligible ChatGPT and Codex text output in the European Union. OpenAI says the rollout covers the EU only and that it is not making text watermarking a global default at launch. OpenAI, BleepingComputer
  • OpenAI says its watermark, called textGrain, works by adjusting the model's word choices and adds no hidden characters, invisible spaces or unusual punctuation. OpenAI, OpenAI Help Center
  • OpenAI says its text watermark links no person, organization, account, prompt or conversation to the text. Anthropic says Claude's watermark carries no identifying information and cannot be traced to a specific person, organization or chat. OpenAI, Anthropic
  • OpenAI said on 5 October 2026 that it was not making its watermark detector public at launch, with access initially limited to approved researchers and expert organizations, case by case. Its help center says checking student essays or theses for AI use is "currently outside the program scope". OpenAI Help Center, OpenAI
  • Anthropic's page on Claude's watermark, updated on 1 September 2026, says its detection API is in private preview for eligible organizations such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, and that it plans to expand access. Anthropic
  • Anthropic's post explaining Claude's text watermark is dated 14 August 2026 and says "Future Claude models will generate text that contains a watermark". Anthropic says it is applying watermarking globally at launch because it does not yet have a durable way to limit it by region. How Claude's text watermark works, How Claude marks AI-generated content
  • OpenAI reported that in an evaluation of 400-token passages, replacing 10% of words with synonyms reduced watermark detection from about 92% to 66%, and replacing 25% reduced it to 17%. OpenAI
  • OpenAI states that the "absence of a detected watermark does not prove human authorship." Its help center adds that the EU AI Act Code of Practice on the Transparency of AI-Generated Content does not require watermarks in outputs shorter than 200 tokens, about 150 words in English. OpenAI, OpenAI Help Center
  • Seven widely used AI detectors labeled human-written TOEFL essays by non-native English speakers as AI-generated 61% of the time on average, while almost never making that error on native speakers' essays, a Stanford study found, as The Markup and Tech & Learning reported. The study did not test Turnitin. The Markup, Tech & Learning
  • Anthropic says light editing probably will not remove Claude's watermark completely, while a complete rewrite will; that when Claude only proofreads a person's text there is little or nothing for the watermark to attach to; and that a translation produced by Claude carries a watermark. Anthropic

Sources

  1. Our approach to EU text provenance rules OpenAI
  2. Provenance signals in OpenAI-generated content OpenAI Help Center
  3. textGrain: Entropy-Calibrated Watermarking for Language Model Text OpenAI
  4. How Claude's text watermark works Anthropic
  5. How Claude marks AI-generated content Anthropic
  6. Article 50: Transparency obligations for providers and deployers of certain AI systems European Commission AI Act Service Desk
  7. AI Detection Tools Falsely Accuse International Students of Cheating The Markup
  8. AI Detectors Discriminate Against Non-Native Speakers, Says Stanford Research Tech & Learning
  9. OpenAI will watermark ChatGPT in the EU but leaves the API opt-in ActuIA
  10. OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU BleepingComputer
  11. ChatGPT text is about to become detectable with invisible watermark to comply with EU AI law Dexerto
  12. Is ChatGPT Watermarking Your Writing? What textGrain Does FindSkill.ai

More in AI news explained